AH-UNet: An Efficient and Enhanced Network for Accurate Melanoma Segmentation

Le résumé fourni par la source

Melanoma (MM) is a highly metastatic and deadly skin disease. However, current diagnostic methods for melanoma face challenges such as invasiveness, poor repeatability, and subjectivity. Therefore, researchers are attempting to use computer technology and algorithms to assist doctors in the diagnosis and treatment of melanoma. In this study, two modules, the adaptive context aware feature attention (ACAFA) and the hybrid attention block with dilated and external mechanisms (HADEM), are first constructed to balance the extraction of global features and local detail features, effectively capturing the multiscale features and long-range dependency relationships of melanoma images. Subsequently, these two modules are combined with the UNet model to propose an optimization network called AH-UNet aimed at improving the accuracy of melanoma segmentation. The feasibility and performance of the AH-UNet model are demonstrated and quantified using the International Skin Imaging Collaboration (ISIC) 2017 and 2018 datasets. Quantitative and qualitative analysis results demonstrate the advantages of the $\mathbf{A H}$-UNet algorithm in handling image noise and skin lesion edge details. Additionally, the model was compared to 8 different deep learning models, achieving optimal segmentation accuracy values of 95.94% and 94.56% in the ISIC-2017 and ISIC-2018 databases respectively, proving the effectiveness of the AH-UNet model in melanoma lesion segmentation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.